Scope, boundaries, and what counts as success

Hardware domains covered

• Actuation: wheel drive (BLDC/PMSM), suction fan, brush/side-brush motors, current sensing and fault handling.
• Perception: ToF/LiDAR, IMU, camera (MIPI-CSI), cliff sensors, bumper switches; power integrity and timing hygiene.
• Compute: MCU/SoC split, memory/storage, reset/watchdog and brownout resilience.
• Energy: battery pack + BMS/fuel gauge, dock contacts, charger behavior, safety interlocks.
• Reliability/EMC: ESD entry points, EMI coupling from motors and DC/DC, dusty/wet field constraints.

Measurable success metrics

“Success” is defined by measurable outcomes and repeatable evidence, not by subjective “it feels better” checks. The following metrics anchor validation and field debug:

Evidence bundle expected for fast diagnosis

• Power: VBAT droop during fan ramp + wheel acceleration; reset reason; charger/BMS fault flags.
• Motion: phase current snapshot + fault pins; encoder ticks vs command; stall counters.
• Sensors: ToF confidence/status, IMU saturation flags, camera frame-drop/error counters, cliff sensor raw + ambient.
• Thermal: NTC profile (pack + driver hotspot proxy), throttling/derating flags, sustained load temperature rise.

Reference architecture: functional blocks + noisy/quiet partitioning

Why the partition model matters

Robot vacuums combine high dI/dt motor switching, sensitive perception sensors, and bursty edge compute on a shared battery. Most “mysterious” field failures are explainable once power loops, return paths, and sensor timing are mapped onto a single reference architecture.

Partition rules (practical, layout-driven)

• Keep motor power loops local: driver → MOSFETs → phase path → shunt return should not share the same return reference as IMU/ToF/camera.
• Define a “quiet sensor island”: sensors + local regulators + reference ground; connect to system ground at a controlled point.
• Place DC/DC by noise class: motor/5V buck(s) near energy entry and drivers; low-noise LDOs near IMU/ToF/camera rails.
• Protect exposed entry points: dock contacts, buttons, bumper metal, and external seams must route ESD away from sensor references.
• Make brownout behavior deterministic: supervisor + reset sequencing should prevent partial-rail states that corrupt sensors or storage.

Interface map (what connects to what)

Design freeze checklist (expensive-to-change items)

• Current sensing topology and headroom: shunt/inline placement, gain/ADC range, and OCP blanking behavior.
• Noise zoning and return paths: motor loop vs sensor island; placement of star point and decoupling strategy.
• Sensor mechanical coupling: IMU mounting stiffness and vibration path; optical window contamination tolerance for ToF/LiDAR.
• MIPI power integrity: camera rails and clock stability under fan/wheel load transients.
• Dock contact protection: ESD path, reverse protection, inrush control, and wet/dust contact resilience.
• Brownout determinism: reset sequencing and “last-gasp” logging so failures are diagnosable.
• Fault observability: expose driver/charger/BMS status bits and counters; ensure logs survive sudden power loss.

Battery pack, BMS, fuel gauge, and the power tree

Field symptoms such as random resets, won’t-charge events, and short runtime are frequently triggered by the power-path and pack protection chain rather than the application logic. A robot vacuum merges bursty motor load steps with sensitive sensors and edge compute on a shared battery—making transient behavior and protection thresholds decisive.

Pack fundamentals (1S/2S/3S concepts) and protection reality

• Cell stack choice drives headroom: higher series count reduces current for the same power but increases protection complexity and balancing needs.
• Protection FETs define “sudden shutoff” behavior: OCP/short protection can disconnect the pack within microseconds–milliseconds, often appearing as an “unexplained reboot.”
• Balancing (multi-cell) is a reliability lever: cell mismatch can cause early UV trips under motor spikes even when average SOC looks healthy.
• NTC placement is not cosmetic: poor thermal coupling (too close to a hot spot or too far from cells) can cause false OTP, low-temp charge refusal, or aggressive derating.

Fuel gauge selection axes (what matters in a vacuum platform)

• Impedance tracking: better at dynamic loads and temperature swings; helps interpret sagging VBAT and aging cells.
• Coulomb counting: simpler but depends on stable calibration points; drift is amplified by temperature and load pulses.
• Learning requirements: a gauge that needs periodic full cycles can report “optimistic SOC” if the usage pattern never reaches learning endpoints.
• Low-temperature behavior: internal resistance rises; a platform may hit UV/protection earlier even with “reasonable” SOC. Gauge algorithms and NTC placement both matter.

Power-tree sequencing: compute vs sensors vs motors

A robust power tree separates noise classes and makes brownout behavior deterministic. Compute rails and storage are sensitive to partial-rail states; sensor rails need low noise; motor rails need surge capability without polluting the quiet domain.

Evidence to capture (fastest path to root cause)

• Correlate time: align VBAT droop, protection flags, and reset reason to a single event timeline.
• Separate “energy” vs “logic”: if resets align with sag/flags, prioritize power-path and protection before software hypotheses.
• Validate temperature gating: charge refusal or early termination often matches NTC thresholds more than charger “fault.”

Wheel drive motors: BLDC/PMSM driver, sensing, and stall detection

Wheel traction and torque delivery are the primary field reliability axis. Carpet friction, thresholds, hair ingestion, and wheel slip create rapid load changes that stress drivers, current sensing, and protection tuning. A resilient design distinguishes “high load” from “fault” using evidence that combines current, speed, and driver state.

Driver choices: integrated driver vs gate driver + MOSFETs

• Integrated drivers: compact and simpler bring-up; limited thermal headroom and fewer layout degrees of freedom.
• Gate driver + MOSFETs: scalable efficiency and current; requires careful loop control, gate ringing management, and EMI containment.
• Headroom matters: margin for peak phase current and VBAT dips prevents false trips and weak torque under real floors.

Current sensing topologies (noise vs accuracy vs protection speed)

Stall detection (evidence-based, algorithm-light)

Robust stall detection avoids single-signal decisions. It combines multiple evidence sources so carpet load is not misclassified as a fault and real jams are not missed.

• Current evidence: sustained overcurrent or repeated spikes beyond a tuned window.
• Speed evidence: encoder/estimated speed deviates from command beyond a threshold (torque not producing motion).
• Driver state evidence: back-EMF/commutation flags, phase open/short indicators, or fault-pin assertions.

Protection pitfalls: shoot-through, deadtime, and OCP blanking

• Shoot-through prevention: inadequate deadtime or gate ringing can cause sudden heating and hard failures.
• Deadtime tradeoff: too short risks cross-conduction; too long reduces effective torque and increases ripple.
• OCP blanking pitfalls: too small causes false trips during start/threshold events; too large delays real short protection.
• Make faults observable: fault pins and status registers should be logged with timestamps for field correlation.

Evidence to capture (what proves the root cause)

• Differentiate load vs fault: high current with matching speed is often “hard floor load”; high current with speed collapse indicates stall/jam.
• Correlate with power: VBAT droop that aligns with stall attempts can trigger system resets—link this to H2-3 evidence.
• Use driver truth: fault pins and commutation flags reduce guesswork and prevent blind parameter changes.

Suction fan + brush/side-brush motors: high RPM, acoustics, and EMI reality

Audible whine, abnormal noise, overheating, and unexpected shutoffs frequently originate from the fan and brush motor domain. These actuators combine fast load transients with long leads and high dV/dt switching—making acoustics, thermal rise, and EMI coupling tightly linked.

High-RPM suction fan motor (often BLDC): where the noise comes from

• Commutation ripple at high RPM: torque ripple and switching edges can excite structural resonances and create a stable tonal peak.
• PWM frequency vs audible band: if PWM or its strong harmonics land inside the audible range, “whine” becomes load-dependent and repeatable.
• Bearing and aero load: bearing wear, dust, and airflow restrictions increase load, shifting both current draw and the dominant acoustic peak.

Brush and side-brush motors: hair ingestion → stall profile → protection tuning

• Progressive drag vs sudden jam: hair wraps often create a rising-torque ramp before full stall; this demands different thresholds than an abrupt obstacle impact.
• Protection tuning tradeoff: aggressive OCP/timeout stops nuisance heating but can misclassify heavy carpet load as a fault.
• Observable recovery: an effective strategy logs stall attempts and recovery success/failure, so field data distinguishes “load” from “fault.”

EMI coupling: motor leads act like antennas (real-world constraints)

Motor phase/lead wiring forms a large loop area and couples energy into sensor interfaces and rails. The goal is not textbook perfection; it is predictable behavior across floors, dust states, and battery voltage.

Evidence to capture (make the problem measurable)

• Thermal slope: temperature rise rate separates friction/ingestion from normal load increase.
• Acoustic fingerprint: track the dominant peak as duty/RPM changes; stable tones are diagnostic.
• Correlation beats intuition: if sensor errors spike with fan PWM transitions, prioritize edge/EMI controls before firmware hypotheses.

SLAM sensor stack hardware: ToF/LiDAR, IMU, camera, cliff & bumper sensors

This section stays hardware-specific: interfaces, noise susceptibility, calibration hooks, and failure signatures. The fastest debug path is to read sensor status and integrity counters, then correlate those signatures with power noise and EMI conditions.

ToF / LiDAR: supply noise sensitivity and optical reality

• Supply noise sensitivity: ripple on emitter/receiver rails can reduce confidence and increase frame errors during motor transients.
• Optical window contamination: dust and smears shift return strength and confidence without changing digital link health.
• Emitter aging: reduced optical power compresses usable range and increases low-confidence events under the same floor and lighting.
• Frame integrity flags: CRC/error counters and validity flags often reveal whether the issue is link/power or optics.

IMU: placement, vibration coupling, and temperature drift

• Placement matters: mounting near fan or wheel vibration paths increases noise and can trigger saturation events during impacts.
• Vibration coupling: mechanical resonances can look like “drift” unless saturation/overflow flags are checked.
• Temperature drift: bias shifts with temperature; calibration hooks should include temperature-aware offsets and logged health states.
• “Bad IMU” signatures: saturation flags, abnormal variance bursts, or repeated re-initialization events in logs.

Camera: MIPI integrity and flicker/rolling artifacts (hardware causes only)

• MIPI-CSI integrity: lane/frame errors or link resets often correlate with EMI bursts and ground/reference disturbances.
• Exposure flicker coupling: LED PWM or motor PWM can couple into rails and manifest as periodic exposure instability.
• Rolling artifacts: supply ripple, reference noise, or timing disturbances can create banding-like effects independent of SLAM processing.

Cliff and bumper sensors: false positives and hardware-level mitigation

• IR cross-talk: emitter/receiver geometry, shielding, and aperture control determine how much self-reflection leaks into measurements.
• Floor reflectivity sensitivity: gloss vs carpet changes return strength; hardware should preserve margin via clean rails and stable mechanical height.
• Ambient immunity: supply isolation and mechanical light blocking reduce false triggers before any algorithmic filtering.

Evidence to capture (turn “sensor weirdness” into measurable signatures)

• CRC up, confidence stable: suspect link integrity or EMI more than optics.
• Confidence down, CRC clean: suspect optical window, emitter aging, or mechanical occlusion before interface tuning.
• Saturation flags during impacts: prioritize mounting/vibration paths and rail cleanliness before calibration changes.

Sensor timing, synchronization, and calibration in production

“Stable on bench but drifting in the field” is frequently explained by timing hygiene and calibration lifecycle control. The objective is deterministic timestamps, bounded latency, and calibration assets that are stored, versioned, and locked for traceability.

Timestamp sources and drift contributors (what breaks determinism)

Calibration assets (what must be stored, versioned, and traceable)

• IMU temperature calibration: bias/scale compensation requires a temperature reference and a stable version tag.
• ToF offset and quality baselines: offsets and confidence baselines must be associated with the sensor module identity.
• Camera intrinsics (storage only): intrinsic parameters should be persisted as immutable assets for that module build.
• Asset packaging: store calibration blocks with CRC, schema ID, and version to prevent silent mismatches.

End-of-line (EOL) flow: calibrate, verify, then lock

• Calibrate: generate IMU temp points, ToF offset checks, and camera intrinsic bundle (if applicable).
• Verify: re-read sensors, check frame integrity counters, and validate that assets load and apply without errors.
• Lock: seal calibration blocks with CRC/version; record a build fingerprint so field logs can detect mismatches.

Evidence to capture (production + field debug signals)

• Jitter distribution: track p50/p95/p99 of timestamp jitter; rising tails indicate contention or EMI-driven ISR disruption.
• Integrity counters: frame CRC and drop counters separate “data quality loss” from “calibration drift.”
• Version mismatch: explicit mismatch logs prevent silent drift caused by wrong calibration blobs.

Edge compute: SoC/MCU split, memory, thermals, and brownout resilience

Compute instability typically presents as reboot, freeze, or navigation jitter. The foundation is a clean partition between real-time control and higher-level compute, plus storage/logging and thermal/brownout behavior that remains predictable under motor transients.

Partitioning: real-time motor MCU vs application/AI SoC

• MCU domain: deterministic motor control, safety interlocks, and time-critical IO where bounded latency is required.
• SoC domain: sensor processing and higher-level compute that benefits from larger memory and throughput.
• Isolation benefit: if the SoC stalls or throttles, the MCU can keep actuators in a safe state and preserve minimal telemetry.

Memory & storage: DDR integrity basics and power-loss-safe logging

• DDR behavior: training and marginal signal integrity can fail at temperature corners, showing up as intermittent crashes or corrupted buffers.
• ECC vs non-ECC: if ECC is not available, error detection via watchdog resets and integrity counters becomes more critical.
• eMMC wear: sustained logging can accelerate wear; reserve budgeted write patterns and store a compact “event ring.”
• Power-loss-safe strategy: use short atomic records and a journaled layout so a brownout does not destroy the last useful evidence.

Thermal: hotspot mapping and load-aware derating

• Hotspot mapping: identify SoC, PMIC, and memory hotspots; temperature gradients matter more than one average reading.
• Derating policy: tie throttling to sensor and motor load so performance reduction is controlled rather than chaotic.
• Telemetry hooks: expose thermal throttle flags so “jitter” can be correlated to thermal events.

Brownout hardening: reset sequencing, supervisors, watchdogs, last-gasp logs

• Reset sequencing: rail order and reset release timing should prevent partial-rail states that corrupt memory or sensors.
• Supervisor + watchdog: use both: supervisor for electrical truth, watchdog for software liveness.
• Last-gasp logging: capture a minimal record (reset reason, rail snapshot, key counters) before rails collapse.

Evidence to capture (turn crashes into classified events)

• Electrical vs software: supervisor brownout flags separate rail collapse from firmware deadlock.
• Thermal correlation: throttle flags aligned with jitter/freeze prevent chasing phantom “navigation bugs.”
• Storage survivability: confirm the last-gasp record remains readable after repeated power events.

Docking & charging subsystem: contacts, inrush, charger IC, and safety

Most “won’t dock / docks but won’t charge / intermittent charge / overheating” complaints can be diagnosed by separating four layers: contact physics, hot-plug/inrush behavior, charger state gates, and safety interlocks. The goal is to turn each symptom into a measurable signature using voltage/current traces, contact resistance trends, and charger status bits.

Dock power chain: where energy is allowed to flow

• Dock contacts establish the supply path, but electrical continuity can be intermittent even when mechanical docking looks fine.
• Reverse protection prevents backfeed and wrong-polarity events; many “connected but no charge” cases are a protection gate not releasing.
• Soft-start / inrush control avoids collapsing the dock supply when the robot hot-plugs a large input capacitance.
• Charger IC applies temperature qualification, safety timers, and termination logic before it commits to fast charging.

Contact physics: pogo-pin wear, oxidation, and misalignment

• Wear / spring fatigue: reduces normal force; micro-open events appear as brief current dropouts and local heating.
• Oxidation / contamination: raises contact resistance (Rc), often causing “charges only if pressed” or “heats at the dock.”
• Misalignment: partial contact can pass “voltage present” checks but fails at higher charge current.
• Measurement method: focus on Rc trend over repeated dock cycles; trending Rc is more diagnostic than a single point.

Inrush & hot-plug: why the first 200 ms matters

• Soft-start target: limit current surge so dock voltage does not collapse during plug-in, especially when the robot’s system rail capacitance is large.
• Reverse protection realism: ideal diode / FET reverse blocks can add drop and gating behavior; validate their turn-on conditions and fault recovery.
• Transient suppression: the dock interface is an ESD/transient entry path; suppression must clamp with a short return path to be effective.
• Common pitfall: an overly slow ramp can trigger charger timeouts or detection windows; an overly fast ramp can trigger brownouts and false faults.

Charger behavior (conceptual gates): phases, NTC, termination, safety timers

Safety interlocks: wet/dust signals and pack fault handling

• Wet/dust detection (if present): treat as a hard gate; log an explicit reason code so “won’t charge” is not ambiguous.
• Pack fault path: charger must honor BMS protection trips and recover cleanly without repeated hot-plug stress cycles.
• Thermal interlock: docking heat can be dominated by contact resistance rather than pack temperature—track both NTC and contact heating signature.

Evidence bundle: a minimal set that closes the loop

• Dock voltage/current: capture the first 200 ms (hot-plug transient) and the steady phase (charge stability).
• Rc trend: log contact resistance proxies (voltage drop at known current) across docking cycles.
• Status bits: record at least one snapshot at “docked, not charging,” and one during “charging steady.”

EMC/ESD/reliability in a dusty, high-dI/dt moving platform

A robot vacuum can pass lab checks but still fail intermittently at home because the field environment combines more entry points (dock, buttons, bumper, exposed metal), stronger noise sources (motor PWM edges, DC/DC switch nodes, long wiring), and constantly changing coupling conditions (movement, dust, humidity, floor reflectivity, dock placement).

ESD entry points: where charge is injected in real homes

• Dock contacts: exposed interface, frequent touch, and hot-plug makes it an ESD/transient gateway.
• Buttons & bumper: high-touch surfaces; a single hit can cause resets, bus lockups, or sensor hangs.
• Exposed metal / trims: offers a discharge point with short path into ground reference disturbances.

EMI sources: who creates the disturbance

• Motor PWM edges: fast current steps and large loops radiate and inject noise back onto rails.
• DC/DC switch nodes: high dV/dt switching nodes couple into nearby sensitive traces and reference planes.
• Long sensor cables: act as antennas; routing and return paths determine whether they inject or receive noise.

Mitigations that survive reality: zoning, return paths, TVS placement, ferrites

• Layout zoning: keep a clear Noisy Zone (motors/DC-DC) separated from the Quiet Sensor Island.
• Return paths: short, continuous return loops often outperform adding components in the wrong location.
• TVS placement logic: clamp at the entry point with a short return path; a TVS far from the connector often fails to protect the victim.
• Ferrites (where they help): most useful on long cables or specific sensitive branches; treat them as targeted impedance, not a universal cure.

Evidence to capture: turn “random” into classified events

Fail reproduction matrix (minimal but effective)

Validation & production test plan (bring-up → EOL → reliability)

This section converts the hardware architecture into an executable plan: a bring-up checklist, a stress/reliability suite that mimics home conditions, and an EOL gate that locks calibration assets and measurable accept windows.

Bring-up checklist (power → motors → sensors → thermal sanity)

Stress & reliability tests (home conditions, made repeatable)

EOL (end-of-line) tests: calibrations, signatures, and gates

• Sensor calibration assets: store offset/version/CRC for IMU temperature calibration and ToF offset; lock these to a build ID.
• Motor current signature gate: check no-load current windows for wheel/fan/brush; deviations often indicate friction, bearing load, or harness resistance.
• Dock contact resistance gate: compute Rc proxy from known current and contact drop; trend it across multiple dock engages.
• Pack health snapshot: record pack voltage, temperature, protection flags, and gauge SoH estimate (if available).

Pass/fail thresholds (measurable accept windows)

BOM/MPN starter list (common choices seen in robot vacuum subsystems)

MPNs are examples for orientation and debugging discussions; selection must match pack voltage, currents, thermal limits, and EMC constraints.

Field debug playbook: symptom → first 3 checks → evidence bundle

This playbook is designed for the first hour of troubleshooting: classify the symptom into the most likely hardware domains, run the first three measurements, and collect a minimal evidence bundle that supports a root-cause decision.

Minimum evidence bundle (reference IDs used in the tables)

Symptom-to-action table (repeatable format)